← All issues

Weekly AI digest

Radiology & medical imaging AI · week of 6 to 12 September 2026

185 peer-reviewed papers · 4 industry and regulatory items · conference and KOL highlights. Adapted from my weekly intelligence report. Full report: FR (PDF) · EN (PDF) · RU (PDF).

Governance takeaway

What to do about it this week

This week hands a radiology leader a clean two-way split in evidence maturity. AI assistance for image interpretation now has trial-grade support: a five-centre randomised crossover in Lancet Digital Health shows an ultrasound aid raising sonographer sensitivity for fetal intracranial malformations by 0.087 while formally protecting specificity. LLM report generation has nothing comparable: a systematic review of 101 studies found not one at low risk of bias and safety data too heterogeneous to pool.

The action: split your AI pipeline into these two evidence classes. For interpretation aids, require reader-performance evidence of the kind that now demonstrably exists. For LLM drafting, treat every pilot as building the safety case from zero, on your own casemix, with your own error counts.

Regulation & AI Act watch

Evidence you can use

Selected from 185 papers indexed in PubMed this week.

Post-market signal

The drift gap between clearance and clinical reality moved into the trade press this week. An Imaging Technology News feature on AI sprawl quotes TestDynamics on the problem its vendor-neutral Satori platform monitors: FDA clearance reflects performance at a point in time, while local populations, scanners and protocols keep changing, so post-deployment drift belongs to the deployer. The metrics named, output drift and radiologist-AI agreement trended over time, are exactly the evidence stream an Article 72 deployment file needs, and declining reader agreement is often the earliest observable signal.

The same week quantified why this matters. A PLOS Digital Health analysis covered by TechTarget found that of more than 1,300 FDA-cleared AI devices, only 2.5 percent were linked to registered prospective trials and only three devices, 0.2 percent, were ever evaluated for patient-centered outcomes. Radiology holds 78 percent of cleared devices and 1 percent of the prospective trials. The burden of demonstrating real-world performance sits almost entirely on deployers, after purchase.

Playbook snippet

One step for your AI committee

Anchor: EU AI Act Article 9 (risk management), pre-deployment.

Step: Before any LLM or classifier pilot, assemble 100 consecutive routine cases from your own service, not teaching cases, and score the candidate tool on them before you see any vendor demo. Record the case list, the model version and the scores in the risk file.

Why: Curated demos systematically overstate: performance on teaching cases does not transfer to routine casemix, and this week's 101-study review shows the published literature cannot carry the safety case for you.

Worked example: Wu M et al. show why the local test matters even for well-built tools: ST-USNet, developed on four institutions with validation AUC 0.984 for malignancy, still performed measurably lower on an independent test cohort, and its authors call for prospective multi-center validation before deployment. Academic Radiology, PMID 42705922, DOI 10.1016/j.acra.2026.08.092.

From the field

The Harvey L. Neiman Health Policy Institute announced a JACR study from Northwell Health: a prospective shadow-mode evaluation of an FDA-cleared aneurysm algorithm across 3,856 CTA examinations. AI added 55 true-positive aneurysms radiologists had not reported, a 39 percent relative detection increase, at the cost of 46 false positives, while radiologists caught 30 aneurysms the AI missed. Performance varied sharply by care setting, favorable inpatient and marginal outpatient, which is the study's real lesson: evaluate AI on your own settings and keep monitoring after go-live.

RSNA published a Board-level update on integrating advanced practice providers into radiologist-led care teams, prioritizing education and convening while deferring standards-setting. The capacity pressure driving APP integration is the same business case behind most imaging AI purchases, which makes the two strategies worth planning together.

On LinkedIn, Bernardo Bizzo of the ACR Data Science Institute announced a new JACR paper describing the LLM engine inside Assess-AI, with a line that frames the season: the first wave of imaging AI produced a flag, the next wave will produce a draft report, and monitoring has to keep up. Amine Korchi dissected Vara's autonomous mammography CE certification and noted he could not find the certified intended-use text, a live demonstration of the first question to ask any autonomous AI vendor. Louis Blankemeier, amplified by Nina Kottler, argued that human and artificial intelligence are non-overlapping designs, the working rationale for complementarity rather than replacement.

Source articles are indexed in PubMed with verified DOIs. Manuscript-stage work is excluded. The full weekly report is produced in French.

Get the weekly digest

One email a week on what matters in radiology and imaging AI. Researched by a radiologist and imaging informatics professional. Free, no vendor sponsorship.